google

/

pix2struct-large

like 26

The image processor of this model has VQA enabled by default:

`Pix2StructProcessor:

• image_processor: Pix2StructImageProcessor {
  "do_convert_rgb": true,
  "do_normalize": true,
  "image_processor_type": "Pix2StructImageProcessor",
  "is_vqa": true,
  "max_patches": 4096,
  "patch_size": {
  "height": 16,
  "width": 16
  },
  "processor_class": "Pix2StructProcessor"
  }

I am having to set it to False before fine-tuning:
processor.image_processor.is_vqa = False

Whereas the base model already has this set to False by default.

Is this expected?